Multi Level Directional Cross Binary Patterns: New handcrafted descriptor for SVM-based texture classification

Abstract

Listen

Translate article

The pattern recognition and computer vision have experienced a prominent progress in feature extraction techniques, judged by the extensive proposed methods in the literature. A big part of these works was devoted to enhance the texture classification performance, regarding the important role of textural analysis in various real-world and challenging applications. Developing discriminant feature extractors requires solid knowledge in machine learning and applied mathematics. However, Local Binary Patterns (LBP) offered much more space to develop enhanced handcrafted descriptors thanks to its simplicity and flexibility. In this paper we introduce a brand new LBP variant referred to as Multi Level Directional Cross Binary Patterns (MLD-CBP). The proposed representation is training-free, low-dimensional, yet discriminative and robust handcrafted operator for texture description. The concept of the proposed MLD-CBP descriptor is based on encoding the most informative directions contained within multi radiuses, which helps in detecting the gray level variations that may occur in different directions. Moreover, the proposed MLD-CBP handcrafted is combined with an automated SVM classifier based on the RBF Kernel, where the γ parameter is calculated automatically according to the training images. Conducted experiments on 15 well known and challenging databases of the literature, demonstrate prominent performance and stability compared to the results achieved by 30 recent and most powerful descriptors of the state-of-the-art. This paper provides also a comparative study on the effect of γ parameter to show the benefits of automatically tuning this parameter value considering the nature of the database and its size.